Qwen-Audio 3.1 cuts voice API prices up to 95%, adds two new models
Alibaba's audio lineup now spans recognition, realtime chat, synthesis and sound design; NVIDIA open-sourced an 8-speaker diarizer the same day.

Copy markdown
Five models, one voice pipeline
ASR, TTS and Realtime all got upgrades, and two new models join the family: ASR-Next adds speaker ID with timestamps plus emotion and ambient-sound detection, while TTS-Next fuses language modeling with diffusion to generate speech, sound effects and background audio in a single pass. The stack now covers understanding, generation, real-time interaction and audio creation end to end.
Voice bills drop by an order of magnitude
Cuts land across the lineup versus the 3.0 series: ASR down as much as 95%, Realtime roughly 85%, TTS about 70% (3.0 TTS Plus was $27.60 per million characters). If you run transcription or voice agents at any volume, the economics just changed.
Realtime that talks and listens at once
The upgraded Realtime model speaks and listens simultaneously with instant interruption, and tunes its pace and empathy to the caller's detected mood — the core loop for a natural phone or voice agent.
The fine print: hosted-only, two models still 'coming soon'
Qwen-Audio 3.x runs through Alibaba Cloud Model Studio as a hosted API, not open weights, and ASR-Next and TTS-Next are marked 'coming soon.' If you need weights you can self-host, the separate Qwen3-TTS repo on GitHub (Apache 2.0) is the option — a different feature set and API.
NVIDIA ships an open-weight 8-speaker diarizer
Nemotron 3 Diarization (100M params, OpenMDW 1.1, commercial use allowed) labels up to 8 overlapping speakers in real time or offline from one checkpoint, posted a 14.72% error rate to top Voice Arena's 12-system board, and pairs with an ASR like Parakeet for speaker-attributed transcripts. Install with `uv pip install 'nemo-toolkit[asr]'`; runs on Ampere through Blackwell GPUs.